The increasing use of social media by retail investors has made online discussions a valuable source of information about market sentiment. However, accurately extracting investors’ opinions from informal financial messages remains challenging, particularly when communication relies on emojis and emotionally charged expressions. Can these elements provide complementary information for identifying investors’ bullish or bearish expectations?
The article Impact of emoji and sentiment-bearing words on financial social-media sentiment classification, by Davide Raffaelli, Daniele Marazzina and Simone Formentin, addresses this question by investigating the predictive contribution of emojis and sentiment-bearing words to financial sentiment classification.
The empirical analysis considers over 1.4 million labeled messages collected from StockTwits between March 2020 and April 2022, focusing on discussions about the SPDR S&P 500 ETF Trust ($SPY). The authors employ a Transformer-based language model, RoBERTa, to classify investors’ self-reported bullish or bearish stance.
A key contribution of the paper is the use of controlled ablation experiments to assess the information conveyed by different components of financial language. Four versions of the messages are considered: original text, text with emoji identities masked, text with sentiment-bearing words masked, and text with both elements masked. This approach makes it possible to evaluate whether emojis provide additional predictive information beyond that contained in explicitly sentiment-related words.
The results reveal substantial differences in classification performance. When models are trained and tested under the same preprocessing conditions, the Matthews Correlation Coefficient (MCC) decreases from 0.7028 for original messages to 0.6506 when emojis are masked and to 0.4745 when sentiment-bearing words are masked. Removing the identity of both types of signals further reduces the MCC to 0.3741. These findings indicate that sentiment-bearing words provide the stronger predictive signal, while emojis convey additional, non-redundant information about investor sentiment.
The study also shows that classification performance generally deteriorates more sharply when information is removed only at the testing stage, highlighting the importance of consistent preprocessing procedures. Additional analyses involving five major U.S. technology stocks confirm that the main findings extend beyond the original $SPY dataset.
Overall, the research highlights the importance of preserving both textual and symbolic information when developing Natural Language Processing tools for financial applications. Emojis should not be regarded merely as decorative elements of online communication: they can provide valuable signals for measuring investor sentiment. These findings have practical implications for the construction of sentiment indicators and the analysis of financial information extracted from social media.
The article is available open access in Finance Research Letters:
https://doi.org/10.1016/j.frl.2026.111057